Demographic-aware temporal graph attention for fair and accurate cardiac abnormality detection in 12-lead ECG

Figures

• 

  Fig. 1. High-Level Architectural Overview of the DA-GAT-v2 Framework.

• 

  Fig. 2. Detailed Architecture of DA-GAT-v2 with FiLM-Based Demographic Conditioning.

• 

  Fig. 3. Overall performance comparison of DA-GAT-v2 against baseline methods on the PTB-XL test set. (a) Macro F1-Score and (b) AUROC. DA-GAT-v2 achieves the highest scores across both metrics, demonstrating superior multi-class ECG classification performance.

• 

  Fig. 4. Per-class F1-Score comparison among DA-GAT-v2, ResNet-50, and DA-GAT-v1 across all five PTB-XL diagnostic categories (NORM, CD, HYP, MI, STTC). DA-GAT-v2 consistently achieves the highest F1-Score in the majority of classes.

• 

  Fig. 5. Fairness evaluation of competing models. (a) Equalized Odds Gap (ΔEO); the dashed red line denotes the clinical acceptability threshold (ΔEO = 0.10), and the shaded region indicates the fair zone. (b) Sex-stratified F1-Score comparing male and female subgroups. DA-GAT-v2 achieves the lowest ΔEO while maintaining balanced performance across sexes.

• 

  Fig. 6. Demographic subgroup analysis evaluating F1-Score across Age × Sex intersectional groups (Young Male, Young Female, Middle-Aged Male, Middle-Aged Female, Elderly Male, Elderly Female). Shaded bands separate age groups. DA-GAT-v2 demonstrates the most consistent performance across all demographic subgroups.

• 

  Fig. 7. Ablation study results across three evaluation metrics: (a) Macro F1-Score, (b) AUROC, and (c) Equalized Odds Gap (ΔEO). Each variant isolates the contribution of a specific model component. The full DA-GAT-v2 model achieves the best trade-off between predictive performance and fairness, with ΔEO below the clinical threshold.

• 

  Fig. 8. Cross-dataset generalization evaluation. (a) Macro F1-Score of each model trained on PTB-XL and evaluated on both PTB-XL and Chapman-Shaoxing datasets. (b) Generalization drop (Δ F1%) represents the performance degradation between the two datasets. DA-GAT-v2 exhibits the smallest cross-dataset performance gap, indicating stronger domain generalization.

• 

  Fig. 9. Demographic-conditioned lead attention weights learned by DA-GAT-v2 across all 12 ECG leads for four demographic groups (Male, Female, Young 18–40 year, Elderly > 65 year). Higher attention values (red) indicate greater model reliance on the corresponding lead for that demographic group, revealing clinically interpretable lead-preference patterns.

• 

  Fig. 10. Training dynamics of DA-GAT-v2 under the three-stage curriculum learning strategy. (a) Training loss, validation loss, and fairness regularization loss across epochs; dashed vertical lines indicate stage transitions (Stage 2: Epoch 15, Stage 3: Epoch 40). (b) Validation Macro F1-Score comparison between DA-GAT-v2, DA-GAT-v1, and Vanilla GAT; the annotated marker indicates the early stopping point.